Presenting a Credit Risk Prediction Model in Financial Institutions Using a Combination of Deep Neural Networks and Big Data Analysis

Fatemeh Ebrahimisarveolya¹

1) PhD Student in Finance and Banking, Faculty of Management and Economics, Azad University, Science and Research Branch, Tehran, Iran

IOI Code : ABFK-ABKHDH
Publication : The third international conference on new researches in management, law, humanities and entrepreneurship(icmlhe.ir)
Abstract :
    Credit risk is one of the most important sources of loss in financial institutions, and the quality of the default probability estimate directly affects the acceptance or rejection of credit applications, facility pricing, credit limit determination, provisioning, and capital management. With the expansion of digital banking, the volume and variety of data that can be used for credit assessment has exceeded traditional financial statement information and repayment records and has also included transactional, behavioral, digital, network, and macroeconomic data. Despite this potential, the use of big data without an appropriate analytical architecture can lead to increased noise, model risk, bias, and difficulty in explainability. The aim of this research is to present a hybrid framework for credit risk prediction in which a big data analytics layer for feature integration, cleaning, and engineering is integrated with a deep neural network and a reinforcement learning branch based on XGBoost, and the output is generated in the form of default probability and risk class that can be used in credit decisions. To evaluate the method, a reproducible experiment was designed on 60,000 simulated credit files with a multi-source structure and a 12-month sequence of repayment behavior. The data was divided into training, validation, and testing with a 60/20/20 ratio, and the logistic regression, random forest, XGBoost, deep neural network, and the proposed hybrid model were compared. The test results showed that the ROC-AUC of the proposed hybrid model is 0.688 and F1 is 0.422, providing a limited but stable improvement in the resolution compared to logistic regression, random forest, and XGBoost. Also, the PR-AUC of the deep neural network reached 0.391, which is a useful indication of the quality of identifying the high-risk class in the unbalanced default problem. The findings show that the main advantage of the proposed framework is not only to increase an accuracy index, but also to create an integrated chain from data to decision that allows for dynamic feature engineering, data drift monitoring, explainability, bias control, and cost-based decision threshold adjustment. Finally, practical implementation requirements including data governance, independent validation, model control, calibration, and continuous monitoring are described.

Keywords : credit risk probability of default deep neural network big data machine learning